pycalphad / pycalphad / 2015

87%

push

travis-ci

FIX: Support PhaseRecord pickling, switch SymEngine backend to LLVM (#264)

This PR includes the following changes:

1. PhaseRecords can now be pickled, provided that all callables are built with LLVM callables. This is from adding internal references (e.g. `ofunc_` to to callable functions that are pickled and reused to create the `FastFunction` objects when unpickling.
2. LLVM callables are now the default everywhere, but this can be configured at runtime or through the `build_functions` or `build_constraint_functions` function APIs.
3. Improved module documentation for `sympydiff_utils` and `build_functions`/`build_constraint_functions`.
4. Updates to pin to `symengine=0.6`. This drastically improves the compile time performance for some of the worst-case systems and makes the compile times go from impossible on LLVM (>20,000 sec) to possible (~30-60 sec).
5. Updates releasing docs and GH PR template to include a checklist item for updating dependencies everywhere (this has been missed in release and tests few times)

Some performance benchmarks for the test suite:

Lambda: 138 s
LLVM `opt_level = 3`: 174 s
LLVM `opt_level=0`: 135 s
LLVM `opt_level=3` everywhere, Lambda Hessians: 125 s

From this data, it does appear that there's some runtime/compile time performance trade off, and the most reasonable defaults is LLVM everywhere (so everything can be pickled).

I'm pretty happy with the design and ability to override the options, but one potential gotcha is that if `build_functions` is called and the result cached, a subsequent change to the `DEFAULT_LAMBDIFY_XYZ` options will not take effect for the cached function (non-cached functions will work as expected). I think this is a pretty small tradeoff for the flexibility of this approach, and the fix for the aforementioned "gotcha" is just to restart the Python kernel and don't compile until the global `DEFAULT_LAMBDIFY_XYZ` is changed. 


* FIX: PhaseRecord: Don't pickle FastFun... (continued)

80 of 80 new or added lines in 3 files covered. (100.0%)

3722 of 4315 relevant lines covered (86.26%)

5.17 hits per line